llama.cpp/examples/quantize/quantize.cpp

151 lines
4.4 KiB
C++
Raw Normal View History

#include "build-info.h"
2023-03-10 18:40:58 +00:00
#include "llama.h"
2023-03-10 18:40:58 +00:00
#include <cstdio>
#include <map>
2023-03-10 18:40:58 +00:00
#include <string>
static const std::map<std::string, llama_ftype> LLAMA_FTYPE_MAP = {
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
{"q4_0", LLAMA_FTYPE_MOSTLY_Q4_0},
{"q4_1", LLAMA_FTYPE_MOSTLY_Q4_1},
{"q5_0", LLAMA_FTYPE_MOSTLY_Q5_0},
{"q5_1", LLAMA_FTYPE_MOSTLY_Q5_1},
{"q8_0", LLAMA_FTYPE_MOSTLY_Q8_0},
{"q2_K", LLAMA_FTYPE_MOSTLY_Q2_K},
{"q3_K", LLAMA_FTYPE_MOSTLY_Q3_K_M},
{"q3_K_S", LLAMA_FTYPE_MOSTLY_Q3_K_S},
{"q3_K_M", LLAMA_FTYPE_MOSTLY_Q3_K_M},
{"q3_K_L", LLAMA_FTYPE_MOSTLY_Q3_K_L},
{"q4_K", LLAMA_FTYPE_MOSTLY_Q4_K_M},
{"q4_K_S", LLAMA_FTYPE_MOSTLY_Q4_K_S},
{"q4_K_M", LLAMA_FTYPE_MOSTLY_Q4_K_M},
{"q5_K", LLAMA_FTYPE_MOSTLY_Q5_K_M},
{"q5_K_S", LLAMA_FTYPE_MOSTLY_Q5_K_S},
{"q5_K_M", LLAMA_FTYPE_MOSTLY_Q5_K_M},
{"q6_K", LLAMA_FTYPE_MOSTLY_Q6_K},
};
bool try_parse_ftype(const std::string & ftype_str, llama_ftype & ftype, std::string & ftype_str_out) {
auto it = LLAMA_FTYPE_MAP.find(ftype_str);
if (it != LLAMA_FTYPE_MAP.end()) {
ftype = it->second;
ftype_str_out = it->first;
return true;
}
// try to parse as an integer
try {
int ftype_int = std::stoi(ftype_str);
for (auto it = LLAMA_FTYPE_MAP.begin(); it != LLAMA_FTYPE_MAP.end(); it++) {
if (it->second == ftype_int) {
ftype = it->second;
ftype_str_out = it->first;
return true;
}
}
}
catch (...) {
// stoi failed
}
return false;
}
2023-03-10 18:40:58 +00:00
// usage:
// ./quantize models/llama/ggml-model.bin [models/llama/ggml-model-quant.bin] type [nthreads]
2023-03-10 18:40:58 +00:00
//
int main(int argc, char ** argv) {
if (argc < 3) {
fprintf(stderr, "usage: %s model-f32.bin [model-quant.bin] type [nthreads]\n", argv[0]);
for (auto it = LLAMA_FTYPE_MAP.begin(); it != LLAMA_FTYPE_MAP.end(); it++) {
fprintf(stderr, " type = \"%s\" or %d\n", it->first.c_str(), it->second);
}
2023-03-10 18:40:58 +00:00
return 1;
}
llama_init_backend();
// parse command line arguments
2023-03-10 18:40:58 +00:00
const std::string fname_inp = argv[1];
std::string fname_out;
int nthread;
llama_ftype ftype;
int arg_idx = 2;
std::string ftype_str;
if (try_parse_ftype(argv[arg_idx], ftype, ftype_str)) {
// argv[2] is the ftype
std::string fpath;
const size_t pos = fname_inp.find_last_of('/');
if (pos != std::string::npos) {
fpath = fname_inp.substr(0, pos + 1);
}
// export as [inp path]/ggml-model-[ftype].bin
fname_out = fpath + "ggml-model-" + ftype_str + ".bin";
arg_idx++;
}
else {
// argv[2] is the output path
fname_out = argv[arg_idx];
arg_idx++;
2023-03-10 18:40:58 +00:00
if (argc <= arg_idx) {
fprintf(stderr, "%s: missing ftype\n", __func__);
return 1;
}
// argv[3] is the ftype
if (!try_parse_ftype(argv[arg_idx], ftype, ftype_str)) {
fprintf(stderr, "%s: invalid ftype '%s'\n", __func__, argv[3]);
return 1;
}
arg_idx++;
}
// parse nthreads
if (argc > arg_idx) {
try {
nthread = std::stoi(argv[arg_idx]);
}
catch (const std::exception & e) {
fprintf(stderr, "%s: invalid nthread '%s' (%s)\n", __func__, argv[arg_idx], e.what());
return 1;
}
} else {
nthread = 0;
}
fprintf(stderr, "%s: build = %d (%s)\n", __func__, BUILD_NUMBER, BUILD_COMMIT);
fprintf(stderr, "%s: quantizing '%s' to '%s' as %s", __func__, fname_inp.c_str(), fname_out.c_str(), ftype_str.c_str());
if (nthread > 0) {
fprintf(stderr, " using %d threads", nthread);
}
fprintf(stderr, "\n");
2023-03-10 18:40:58 +00:00
const int64_t t_main_start_us = llama_time_us();
2023-03-10 18:40:58 +00:00
int64_t t_quantize_us = 0;
// load the model
{
const int64_t t_start_us = llama_time_us();
2023-03-10 18:40:58 +00:00
if (llama_model_quantize(fname_inp.c_str(), fname_out.c_str(), ftype, nthread)) {
2023-03-10 18:40:58 +00:00
fprintf(stderr, "%s: failed to quantize model from '%s'\n", __func__, fname_inp.c_str());
return 1;
}
t_quantize_us = llama_time_us() - t_start_us;
2023-03-10 18:40:58 +00:00
}
// report timing
{
const int64_t t_main_end_us = llama_time_us();
2023-03-10 18:40:58 +00:00
printf("\n");
printf("%s: quantize time = %8.2f ms\n", __func__, t_quantize_us/1000.0);
printf("%s: total time = %8.2f ms\n", __func__, (t_main_end_us - t_main_start_us)/1000.0);
2023-03-10 18:40:58 +00:00
}
return 0;
}